High-precision machine learning identifies a reproducible functional connectivity signature of autism spectrum diagnosis in a subset of individuals

Background Discovery of predictive biomarkers is essential for understanding the neurobiological underpinnings of autism spectrum diagnosis (ASD) and improving identification. Resting-state functional connectivity analyses of individuals with ASD have established sensitivity of brain connectivity at the group level. However, the extensive heterogeneity in ASD limits the translation of these findings into reliable individual-level biomarkers. We analyzed the Autism Brain Imaging Data Exchange 1 and 2 datasets, calculating Pearson’s correlation-based functional connectivity across 18 brain networks. Using transductive conformal prediction, a machine learning approach that assigns confidence scores to predictions based on conformality to known classes, we classified individuals with ASD and neurotypical controls. Results By combining predictors into an ensemble using hierarchical agglomerative clustering, we identified a signature that confers a more than 7-fold increase in individual risk of ASD, yet is still identified in an estimated 1 in 200 individuals in the general population. The individual risk conferred by the model is increased 4-fold over that of previously published imaging models and outperforms the current state of the art in precision for ASD classification. The high-risk signature was characterized by underconnectivity of transmodal brain networks, including the frontoparietal and basal ganglia network, and subcomponents of the limbic and default mode networks. Conclusions A highly targeted prediction model can identify a subset of functional connectivity alterations that confer high risk for ASD at the individual level, which may be masked by traditional machine learning models due to ASD heterogeneity. Results could help disentangle the multitude of etiological pathways and behavioral symptoms that challenge our understanding of ASD by focusing on highly penetrant connectivity signatures.